ORL (On-Policy Residual Learning) is an open-source project released by the Reddit machine learning community, aimed at accelerating the inference of diffusion language models (dLLM), covering model weights, training code, evaluation pipeline and inference engine.